This is the market-facing line of my work, in one place. It starts with ten years of MetaTrader/MQL engineering — Expert Advisors, custom indicators, tick-level backtesting and risk frameworks — and runs through to real-time market-data pipelines that stream live ticks off broker terminals and monitor live crypto venues, score them with disciplined multi-model decision systems, and gate every action behind honest measurement. Built, run, and repaired end-to-end on my own multi-node platform, the Forge. The through-line maps directly onto real-time market-data architecture: ingest, normalize, distribute low-latency, validate rigorously, manage risk, and own it from wire to decision.
Read this first. This is the work of a solo research-and-production lab (NexCore AI), not a record of enterprise client deployments. Every performance figure below is a measured evaluation metric from my own test harness — research instrumentation, not a returns claim and not financial advice. Live trading capital is small and deliberately not overstated; most systems here are read-only market-data monitoring and offline research, with any execution gated behind manual approval. What I am showing is capability, engineering discipline, and depth — a decade of it.
The role is multi-channel real-time market-data infrastructure: low-latency, high-throughput distribution, fail-over and DR, wired into trading systems and market-data flows. Here is the honest mapping — each requirement against concrete, running work on the Forge, followed by the systems themselves.
From the MetaTrader foundation to the real-time crypto era. Expand any card for architecture, data flow, and what's actually measured. Badges say what each one is — a foundation, a live market-data pipeline, or active research.
My market work begins on MetaTrader. Across roughly a decade I wrote Expert Advisors and custom indicators in MQL4 and MQL5, tested and compiled in the terminal, and pursued a deliberately mathematical, non-indicator approach to trading — price structure and geometry rather than lagging oscillators. That philosophy became the “9-point” method: a structured multi-factor score rather than a single crossover signal. I validated in the Strategy Tester on historical data pulled straight from the terminal, down to tick level, with slippage and spread modelled and risk-per-trade and reward-to-risk fixed in code. The result is a large personal body of work — a ~600 MB corpus of MQL (roughly half original, half studied third-party EAs and indicators; one archive alone holds 156 source files, including a 6,000-line multi-type grid EA and polynomial-regression indicators). Two research threads grew out of it: MetaTrader↔Python↔LLM bridges to drive strategies from modern code, and an exploratory LLM-consensus approach to reconstructing compiled .ex4/.ex5 EA logic where conventional decompilers fail.
This is the MetaTrader work turned into a real-time system, and it is my strongest live market-data credential. A bridge Expert Advisor streams live tick and bar data off two real MetaTrader terminals — Admiral (1:30, EU-regulated) and RoboForex — across eight FX symbols, writing through a RAM-drive file channel at roughly 5 ms so the Python side reads a live feed without touching the SSD. On each bar close the 9-point scorer grades the market on three pillars — Structure, Momentum, Exhaustion, three checks each — while a regime detector classifies trending / ranging / volatile / exhausting and a nested genetic-algorithm daemon optimizes parameters on 4-hour and 24-hour cycles. Nothing reaches an order until it clears a 7-gate risk filter, and every strategy is measured on a walk-forward engine reporting Sharpe, profit factor and drawdown. Roughly 17,000 lines across 43 files hold it together.
TradeArchitect adds a decision layer on top of the market-data pipeline. It reasons about a specific cluster of recent M15 candles and makes a binary call — long / short / skip — which a vectorized OHLCV replay oracle validates immediately against what price actually did; each episode is stored as few-shot memory so later calls learn from similar past setups, and an adaptive layer tunes the confidence threshold over time. Models earn their seat the hard way: per-timeframe champions are selected by measured evaluation win rate (the M15 champion measured at 75.8%, a reasoning model at 58.1%), and the top timeframe runs a live three-model consensus — local champion, an API reasoning model, and a frontier judge — that fires only on a 2-of-3 majority. Risk is explicit: fixed reward-to-risk, a break-even move, a mandatory higher-timeframe trend gate, friction budgeting, session filtering, a daily-loss kill and demo-first safety with a manual approval gate. To be precise: these are measured harness metrics, this is research instrumentation, not a returns claim, and the discipline of measuring honestly is the actual product.
The crypto side starts with real-time monitoring. A monitor on the Bybit v5 API continuously mirrors open USDT-perpetual positions exactly as the exchange web panel shows them, with an account-level risk pulse and per-position distance-to-liquidation alerts. On top of it sits factor-neutral liquidation research: a book that neutralizes correlated systematic risk still carries idiosyncratic risk, so the model adds an explicit idiosyncratic buffer and a liquidation-surface control — k_total = (equity − maintenance-margin) / total-PnL-sigma — with a hard kill-switch when the safety distance collapses, and per-leg tail/CVaR guards on the roadmap. The engineering was peer-reviewed by a multi-model panel and passes tiered verification. The track record is kept deliberately honest and small: about +$180 realized on ~$1,000 staked over a year — real, measured, and explicitly not dressed up as anything larger. A separate, exploratory research thread on the same venue applies autoregressive linear-prediction (Burg) signal methods with a multi-source fusion layer to real-time multi-instrument futures; a 16.5-hour instrumented live run surfaced an overtrading failure mode that a stability-focused revision then addressed. This is early-stage signal research — a work in progress, not a validated strategy.
MEXC is the multi-venue proof: the Bybit monitoring discipline, adapted to a second exchange rather than rebuilt. A real-time agent polls positions, unrealized PnL, margin headroom and funding, runs a dead-capital detector for no-stop lottery-ticket positions, and pushes alerts to my messaging channel. I also ran a structured recon of the exchange's own in-app AI, benchmarking it as a zero-weight sentiment feed (it turned out to be an engagement funnel, not an analyst) and verified live-vs-cached data freshness on the tape. Risk guardrails were ratified by my multi-agent deliberation body — account-level liquidation-distance thresholds (warn / escalate) and a dead-capital rule — and any write capability is gated behind a manual Tier-3 approval, reduce-only first, after a paper-replay gate. An API-parity check against Bybit keeps a single normalized model across both venues.
SpeculativeOracle is the forecasting-and-calibration research behind the market work. Given a set of premises it computes conditional outcomes faster than wall-clock — a deterministic simulation lane plus an ensemble-spread lane — and streams them before the operator has to commit; on a chaotic reference system it ran ~805–837× ahead of real time. What makes it useful for markets is the honesty machinery: a calibration immune system scores every prediction by Brier skill against a persistence baseline and retires anything that stops beating it, a basin classifier labels each forecast robust / marginal / knife-edge with an empirical predictability horizon, and a bitemporal decision register keeps full provenance. It was reviewed through three layers — deliberation body, design critique, and peer review — and uses my trading systems strictly as read-only walk-forward calibration test beds. No retrocausality is claimed or implied; for a deterministic system, prediction is just fast retrieval.
The market systems are not islands. They reuse the same foundations I built once and apply everywhere — a consensus engine, an adaptive optimizer, a verification ladder, a disagreement institution, and a shared memory.
The market systems run on a self-hosted, multi-node GPU platform I designed and administer end-to-end. Services run under systemd; public edges go through Cloudflare tunnels (this page among them); a Kubernetes (k3s) cluster carries the observability stack — Grafana, Prometheus and Alertmanager for metrics, Loki + promtail for logs — alongside GPU model pools and vector stores.
Resilience is built in rather than bolted on: node-liveness watchdogs, automated cross-node backups, retry/back-off and reconnect logic in the data clients, and tiered verification gates before anything is promoted. The point is that a real-time data platform is only as good as its fail-over, its observability, and the honesty of its measurements — and those are the parts I care about most.
> 25+ years in enterprise infrastructure and telecom. I run AI and market-data systems the way production is run, because that is where I come from.
AI Systems Engineer — 25+ years across enterprise infrastructure and telecom, most recently as a Principal Technical Support Engineer, and roughly ten years of algorithmic trading in parallel — beginning in 2013 with formal, hands-on trading courses at a licensed forex brokerage in Bucharest, then years of MetaTrader/MQL Expert-Advisor development, and now real-time market-data systems. MSc Open Systems Architecture and BSc Electronics & Computer Science, Politehnica University of Bucharest. Oracle Cloud Infrastructure 2025 Generative AI Professional and AI Foundations Associate.
Everything on this page is real software I designed, built, and operate solo — a decade of demonstrated capability and depth in trading systems and real-time market-data flows, offered honestly as exactly that. Open to a Real-Time Market Data / Solutions Architect engagement — B2B contract or full-time.
Email is the front door — or use the form on the main site. Prefer voice? Request a call by email and I'll confirm a slot.
./email me ./contact form ./linkedin ./full portfolio